Conclusion: before scoring an agent's latency or tool-call failures as a policy regression, prove the kernel is not reclaim-throttling the process. Under cgroup v2, memory.high does not kill anything; it applies direct reclaim pressure in the allocating task's context. The observable symptom looks exactly like a policy failure from the outside: long, uneven stalls inside ordinary syscalls such as write, mmap, and page faults, truncated or dropped tool outputs, and higher tail latency with no error code. Because no OOM event fires, dashboards that only watch oom_kill counts will report the run as clean.

A reproducible diagnostic is to treat the memory.high value as the independent variable while holding policy, seed, and workload fixed. Run one arm with memory.high unset or raised to memory.max, and one constrained arm, then compare the latency distribution rather than the mean. Read memory.events for the high counter and memory.stat for pgscan, pgsteal, workingset_refault, and pgmajfault; sample memory.pressure (some and full) before and after each run so you can see whether the delay coincides with reclaim rather than with tool execution. If the constrained arm's tail grows only in calls that allocate and the pressure counters rise while policy decisions are unchanged, the kernel is the more parsimonious explanation. Cross-check cpu.pressure so CPU scheduling delay is not mislabeled as reclaim, and record the exact cgroup path, because sidecars, log shippers, or a page-cache-heavy file read can be charged to a different cgroup than the agent and leave the agent's own counters looking clean.

Limitations to state in the write-up: PSI requires CONFIG_PSI and is not always exposed inside containers; memory.high reclaim is charged to the allocating cgroup, so page cache from reading large context files can trigger it without the model's working set growing; short-lived tool subprocesses may exit before sampling catches them; and throttling intensity depends on kernel version and cgroup memory controller behavior, so results are not portable across hosts unless kernel and runtime are pinned. Useful reference directions are the kernel cgroup v2 memory documentation covering memory.high semantics, memory.events, and memory.stat fields, plus the PSI interface docs, applied through a two-arm ablation that reports per-call latency with pressure counters attached. If no pressure signal appears in the constrained arm, drop the reclaim hypothesis and attribute the difference elsewhere instead of reporting it as a policy effect.